Text-to-image generation on MS-COCO 2017
26.91FIDDiffSparse
Evaluation results
| Method | Links | ||||||||
|---|---|---|---|---|---|---|---|---|---|
| DiffSparseMACs (T)=1.64, Speedup=1.74×, Ratio (R)=43%, Sampling steps=20, Base model=PixArt-α2026.04 | 26.91 | — | — | — | 0.164 | — | — | — | |
| SODANs=8, Latency (s)=0.472, FLOPs (T)=5.95, Spe.=1.88×2026.03 | 27.33 | — | — | — | 16.42 | — | — | — | |
| ProCache2025.12 | 27.66 | 1.215 | 5.7 | 1.96 | 16.45 | — | — | — | |
| SODANs=7, Latency (s)=0.428, FLOPs (T)=5.07, Spe.=2.21×2026.03 | 27.72 | — | — | — | 16.44 | — | — | — | |
| DiffSparseMACs (T)=1.30, Speedup=1.91×, Ratio (R)=54%, Sampling steps=20, Base model=PixArt-α2026.04 | 27.79 | — | — | — | 0.164 | — | — | — | |
| DuCaMACs (T)=1.63, Speedup=1.78×, Sampling steps=20, Base model=PixArt-α2026.04 | 27.98 | — | — | — | 0.164 | — | — | — | |
| ToCa2025.12 | 28.02 | 1.473 | 6.33 | 1.77 | 16.43 | — | — | — | |
| DuCaLatency (s)=0.479, FLOPs (T)=5.99, Spe.=1.87×2026.03 | 28.05 | — | — | — | 16.42 | — | — | — | |
| PixArt-αSteps=20, Latency (s)=0.857, FLOPs (T)=11.18, Spe.=1.00×2026.03 | 28.1 | — | — | — | 16.29 | — | — | — | |
| PixArt-α2025.12 | 28.12 | 2.142 | 11.18 | 1 | 16.29 | — | — | — | |
| DiCacheMACs (T)=1.63, Speedup=1.77×, Sampling steps=20, Base model=PixArt-α2026.04 | 28.19 | — | — | — | 0.164 | — | — | — | |
| PixArt-αMACs (T)=2.86, Speedup=1.00×, Sampling steps=20, Base model=PixArt-α2026.04 | 28.2 | — | — | — | 0.163 | — | — | — | |
| ToCaMACs (T)=1.64, Speedup=1.75×, Sampling steps=20, Base model=PixArt-α2026.04 | 28.35 | — | — | — | 0.164 | — | — | — | |
| ToCaLatency (s)=0.492, FLOPs (T)=6.05, Spe.=1.85×2026.03 | 28.38 | — | — | — | 16.43 | — | — | — | |
| Δ-DiT2025.12 | 28.91 | 1.724 | 7.68 | 1.54 | 16.41 | — | — | — | |
| TaylorSeerMACs (T)=1.57, Speedup=1.83×, Sampling steps=20, Base model=PixArt-α2026.04 | 29.08 | — | — | — | 0.163 | — | — | — | |
| DeepCacheMACs (T)=1.48, Speedup=1.61×, cache interval N=2, Sampling steps=20, Base model=PixArt-α2026.04 | 29.61 | — | — | — | 0.163 | — | — | — | |
| FORAactivation_interval_N=22025.12 | 29.63 | 1.575 | 5.66 | 1.98 | 16.4 | — | — | — | |
| FORAFLOPs (T)=5.66, Spe.=1.98×2026.03 | 29.67 | — | — | — | 16.4 | — | — | — | |
| FORAMACs (T)=1.43, Speedup=1.64×, cache interval N=2, Sampling steps=20, Base model=PixArt-α2026.04 | 29.67 | — | — | — | 0.164 | — | — | — | |
| FORAactivation_interval_N=32025.12 | 29.84 | 1.301 | 6.05 | 2.79 | 16.42 | — | — | — | |
| LANTERN2026.03 | 30.72 | 33.8 | — | — | 32.7 | 655.37 | 2.35 | 3.65 | |
| MC-SJDConfiguration=MC-SJD (So et al., 2025a)2026.03 | 30.73 | 33.28 | — | — | 31.32 | 814.5 | 2.6 | 2.92 | |
| LANTERN + Ours2026.03 | 30.74 | 31.92 | — | — | 32.72 | 606.28 | 2.71 | 3.92 | |
| SJD-VPConfiguration=SJD-VP2026.03 | 30.77 | 26.95 | — | — | 31.42 | 729.7 | 3.21 | 3.26 | |
| Jacobi Decoding2026.03 | 30.78 | 85.64 | — | — | 31.31 | 2,312 | 1.01 | 1.03 | |
| SJD2026.03 | 30.78 | 40.1 | — | — | 31.31 | 1,058.6 | 2.22 | 2.25 | |
| Jacobi DecodingConfiguration=Jacobi Decoding (Song et al., 2021)2026.03 | 30.78 | 85.64 | — | — | 31.31 | 2,312 | 1.01 | 1.03 | |
| SJDConfiguration=SJD (Teng et al., 2024)2026.03 | 30.78 | 40.1 | — | — | 31.31 | 1,058.6 | 2.16 | 2.25 | |
| Lumina-mGPT2026.03 | 30.79 | 86.55 | — | — | 31.31 | 2,379 | 1 | 1 | |
| Lumina-mGPTConfiguration=Lumina-mGPT (Liu et al., 2024b)2026.03 | 30.79 | 86.55 | — | — | 31.31 | 2,379 | 1 | 1 | |
| SJD + Ours2026.03 | 30.96 | 36.47 | — | — | 31.4 | 945.2 | 2.37 | 2.52 | |
| GSD + OursG=32026.03 | 31.11 | 32.55 | — | — | 31.37 | 803.4 | 2.66 | 2.96 | |
| GSD-VPConfiguration=GSD-VP2026.03 | 31.43 | 30.78 | — | — | 31.46 | 689.58 | 2.81 | 3.45 | |
| GSDG=32026.03 | 31.5 | 34.12 | — | — | 31.33 | 925.89 | 2.54 | 2.56 | |
| GSDConfiguration=GSD (So et al., 2025b)2026.03 | 31.5 | 34.12 | — | — | 31.33 | 925.89 | 2.54 | 2.57 | |
| LANTERN-VPConfiguration=LANTERN-VP2026.03 | 31.7 | 26.71 | — | — | 31.43 | 590.33 | 3.24 | 4.03 | |
| LANTERNConfiguration=LANTERN(Jang et al., 2024b)2026.03 | 31.72 | 33.8 | — | — | 31.32 | 655.37 | 2.56 | 3.63 | |
| PixArt-αSteps=10, Latency (s)=0.491, FLOPs (T)=5.54, Spe.=2.00×2026.03 | 37.57 | — | — | — | 15.83 | — | — | — | |
| 50% stepsMACs (T)=1.43, Speedup=1.74×, Sampling steps=10 (50% of 20), Base model=PixArt-α2026.04 | 37.57 | — | — | — | 0.158 | — | — | — | |
| PixArt-αsampling_steps=50%2025.12 | 37.62 | 1.044 | 5.59 | 2 | 15.81 | — | — | — |